Finding Deceptive Opinion Spam by Any Stretch of the Imagination

7/29/2019 Finding Deceptive Opinion Spam by Any Stretch of the Imagination

1/11

Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics, pages 309319,Portland, Oregon, June 19-24, 2011. c2011 Association for Computational Linguistics

Finding Deceptive Opinion Spam by Any Stretch of the Imagination

Myle Ott Yejin Choi Claire Cardie

Department of Computer ScienceCornell UniversityIthaca, NY 14853

{myleott,ychoi,cardie}@cs.cornell.edu

Jeffrey T. Hancock

Department of CommunicationCornell UniversityIthaca, NY 14853

[email protected]

Abstract

Consumers increasingly rate, review and re-

search products online (Jansen, 2010; Litvin

et al., 2008). Consequently, websites con-

taining consumer reviews are becoming tar-

gets of opinion spam. While recent workhas focused primarily on manually identifi-

able instances of opinion spam, in this work

we study deceptive opinion spamfictitious

opinions that have been deliberately written to

sound authentic. Integrating work from psy-

chology and computational linguistics, we de-

velop and compare three approaches to detect-

ing deceptive opinion spam, and ultimately

develop a classifier that is nearly 90% accurate

on our gold-standard opinion spam dataset.

Based on feature analysis of our learned mod-

els, we additionally make several theoretical

contributions, including revealing a relation-ship between deceptive opinions and imagina-

tive writing.

1 Introduction

With the ever-increasing popularity of review web-

sites that feature user-generated opinions (e.g.,

TripAdvisor1 and Yelp2), there comes an increasing

potential for monetary gain through opinion spam

inappropriate or fraudulent reviews. Opinion spam

can range from annoying self-promotion of an un-

related website or blog to deliberate review fraud,

as in the recent case3 of a Belkin employee who

1http://tripadvisor.com2http://yelp.com

3http://news.cnet.com/8301-1001_

3-10145399-92.html

hired people to write positive reviews for an other-

wise poorly reviewed product.4

While other kinds of spam have received consid-

erable computational attention, regrettably there has

been little work to date (see Section 2) on opinionspam detection. Furthermore, most previous work in

the area has focused on the detection ofDISRUPTIVE

OPINION SPAMuncontroversial instances of spam

that are easily identified by a human reader, e.g., ad-

vertisements, questions, and other irrelevant or non-

opinion text (Jindal and Liu, 2008). And while the

presence of disruptive opinion spam is certainly a

nuisance, the risk it poses to the user is minimal,

since the user can always choose to ignore it.

We focus here on a potentially more insidi-

ous type of opinion spam: DECEPTIVE OPINION

SPAMfictitious opinions that have been deliber-

ately written to sound authentic, in order to deceive

the reader. For example, one of the following two

hotel reviews is truthful and the other is deceptive

opinion spam:

1. I have stayed at many hotels traveling for both business

and pleasure and I can honestly stay that The James is

tops. The service at the hotel is first class. The rooms

are modern and very comfortable. The location is per-

fect within walking distance to all of the great sights and

restaurants. Highly recommend to both business trav-

ellers and couples.

2. My husband and I stayed at the James Chicago Hotel

for our anniversary. This place is fantastic! We knew

as soon as we arrived we made the right choice! The

rooms are BEAUTIFUL and the staff very attentive and

wonderful!! The area of the hotel is great, since I love

to shop I couldnt ask for more!! We will definatly be

4It is also possible for opinion spam to be negative, poten-

tially in order to sully the reputation of a competitor.

309


2/11

back to Chicago and we will for sure be back to the James

Chicago.

Typically, these deceptive opinions are neither

easily ignored nor even identifiable by a human

reader;5 consequently, there are few good sources

of labeled data for this research. Indeed, in the ab-

sence of gold-standard data, related studies (see Sec-tion 2) have been forced to utilize ad hoc procedures

for evaluation. In contrast, one contribution of the

work presented here is the creation of the first large-

scale, publicly available6 dataset for deceptive opin-

ion spam research, containing 400 truthful and 400

gold-standard deceptive reviews.

To obtain a deeper understanding of the nature of

deceptive opinion spam, we explore the relative util-

ity of three potentially complementary framings of

our problem. Specifically, we view the task as: (a)

a standard text categorization task, in which we usen-grambased classifiers to label opinions as either

deceptive or truthful (Joachims, 1998; Sebastiani,

2002); (b) an instance of psycholinguistic decep-

tion detection, in which we expect deceptive state-

ments to exemplify the psychological effects of ly-

ing, such as increased negative emotion and psycho-

logical distancing (Hancock et al., 2008; Newman et

al., 2003); and, (c) a problem ofgenre identification,

in which we view deceptive and truthful writing as

sub-genres of imaginative and informative writing,

respectively (Biber et al., 1999; Rayson et al., 2001).

We compare the performance of each approach

on our novel dataset. Particularly, we find that ma-

chine learning classifiers trained on features tradi-

tionally employed in (a) psychological studies of

deception and (b) genre identification are both out-

performed at statistically significant levels by n-

grambased text categorization techniques. Notably,

a combined classifier with both n-gram and psy-

chological deception features achieves nearly 90%

cross-validated accuracy on this task. In contrast,

we find deceptive opinion spam detection to be well

beyond the capabilities of most human judges, who

perform roughly at-chancea finding that is consis-

tent with decades of traditional deception detection

research (Bond and DePaulo, 2006).

5The second example review is deceptive opinion spam.6Available by request at: http://www.cs.cornell.

edu/myleott/op_spam

Additionally, we make several theoretical con-

tributions based on an examination of the feature

weights learned by our machine learning classifiers.

Specifically, we shed light on an ongoing debate in

the deception literature regarding the importance of

considering the context and motivation of a decep-

tion, rather than simply identifying a universal set

of deception cues. We also present findings that are

consistent with recent work highlighting the difficul-

ties that liars have encoding spatial information (Vrij

et al., 2009). Lastly, our study of deceptive opinion

spam detection as a genre identification problem re-

veals relationships between deceptive opinions and

imaginative writing, and between truthful opinions

and informative writing.

The rest of this paper is organized as follows: in

Section 2, we summarize related work; in Section 3,

we explain our methodology for gathering data and

evaluate human performance; in Section 4, we de-scribe the features and classifiers employed by our

three automated detection approaches; in Section 5,

we present and discuss experimental results; finally,

conclusions and directions for future work are given

in Section 6.

2 Related Work

Spam has historically been studied in the contexts of

e-mail (Drucker et al., 2002), and the Web (Gyongyi

et al., 2004; Ntoulas et al., 2006). Recently, re-

searchers have began to look at opinion spam aswell (Jindal and Liu, 2008; Wu et al., 2010; Yoo

and Gretzel, 2009).

Jindal and Liu (2008) find that opinion spam is

both widespread and different in nature from either

e-mail or Web spam. Using product review data,

and in the absence of gold-standard deceptive opin-

ions, they train models using features based on the

review text, reviewer, and product, to distinguish

between duplicate opinions7 (considered deceptive

spam) and non-duplicate opinions (considered truth-

ful). Wu et al. (2010) propose an alternative strategyfor detecting deceptive opinion spam in the absence

7Duplicate (or near-duplicate) opinions are opinions that ap-

pear more than once in the corpus with the same (or similar)

text. While these opinions are likely to be deceptive, they are

unlikely to be representative of deceptive opinion spam in gen-

eral. Moreover, they are potentially detectable via off-the-shelf

plagiarism detection software.

310


3/11

of gold-standard data, based on the distortion of pop-

ularity rankings. Both of these heuristic evaluation

approaches are unnecessary in our work, since we

compare gold-standard deceptive and truthful opin-

ions.

Yoo and Gretzel (2009) gather 40 truthful and 42

deceptive hotel reviews and, using a standard statis-

tical test, manually compare the psychologically rel-

evant linguistic differences between them. In con-

trast, we create a much larger dataset of 800 opin-

ions that we use to develop and evaluate automated

deception classifiers.

Research has also been conducted on the re-

lated task of psycholinguistic deception detection.

Newman et al. (2003), and later Mihalcea and

Strapparava (2009), ask participants to give both

their true and untrue views on personal issues

(e.g., their stance on the death penalty). Zhou et

al. (2004; 2008) consider computer-mediated decep-tion in role-playing games designed to be played

over instant messaging and e-mail. However, while

these studies compare n-grambased deception clas-

sifiers to a random guess baseline of 50%, we addi-

tionally evaluate and compare two other computa-

tional approaches (described in Section 4), as well

as the performance of human judges (described in

Section 3.3).

Lastly, automatic approaches to determining re-

view quality have been studieddirectly (Weimer

et al., 2007), and in the contexts of helpful-ness (Danescu-Niculescu-Mizil et al., 2009; Kim et

al., 2006; OMahony and Smyth, 2009) and credibil-

ity (Weerkamp and De Rijke, 2008). Unfortunately,

most measures of quality employed in those works

are based exclusively on human judgments, which

we find in Section 3 to be poorly calibrated to de-

tecting deceptive opinion spam.

3 Dataset Construction and Human

Performance

While truthful opinions are ubiquitous online, de-

ceptive opinions are difficult to obtain without re-

sorting to heuristic methods (Jindal and Liu, 2008;

Wu et al., 2010). In this section, we report our ef-

forts to gather (and validate with human judgments)

the first publicly available opinion spam dataset with

gold-standard deceptive opinions.

Following the work of Yoo and Gretzel (2009), we

compare truthful and deceptive positive reviews for

hotels found on TripAdvisor. Specifically, we mine

all 5-star truthful reviews from the 20 most popular

hotels on TripAdvisor8 in the Chicago area.9 De-

ceptive opinions are gathered for those same 20 ho-

tels using Amazon Mechanical Turk10 (AMT). Be-

low, we provide details of the collection methodolo-

gies for deceptive (Section 3.1) and truthful opinions

(Section 3.2). Ultimately, we collect 20 truthful and

20 deceptive opinions for each of the 20 chosen ho-

tels (800 opinions total).

3.1 Deceptive opinions via Mechanical Turk

Crowdsourcing services such as AMT have made

large-scale data annotation and collection efforts fi-

nancially affordable by granting anyone with ba-

sic programming skills access to a marketplace of

anonymous online workers (known as Turkers) will-ing to complete small tasks.

To solicit gold-standard deceptive opinion spam

using AMT, we create a pool of 400 Human-

Intelligence Tasks (HITs) and allocate them evenly

across our 20 chosen hotels. To ensure that opin-

ions are written by unique authors, we allow only a

single submission per Turker. We also restrict our

task to Turkers who are located in the United States,

and who maintain an approval rating of at least 90%.

Turkers are allowed a maximum of 30 minutes to

work on the HIT, and are paid one US dollar for anaccepted submission.

Each HIT presents the Turker with the name and

website of a hotel. The HIT instructions ask the

Turker to assume that they work for the hotels mar-

keting department, and to pretend that their boss

wants them to write a fake review (as if they were

a customer) to be posted on a travel review website;

additionally, the review needs to sound realistic and

portray the hotel in a positive light. A disclaimer

8TripAdvisor utilizes a proprietary ranking system to assess

hotel popularity. We chose the 20 hotels with the greatest num-ber of reviews, irrespective of the TripAdvisor ranking.

9It has been hypothesized that popular offerings are less

likely to become targets of deceptive opinion spam, since the

relative impact of the spam in such cases is small (Jindal and

Liu, 2008; Lim et al., 2010). By considering only the most

popular hotels, we hope to minimize the risk of mining opinion

spam and labeling it as truthful.10http://mturk.com

311


4/11

Time spent t (minutes)

All submissions

count: 400

tmin: 0.08, tmax: 29.78

t: 8.06, s: 6.32

Length (words)

All submissionsmin: 25, max: 425

: 115.75, s: 61.30

Time spent t < 1

count: 47

min: 39, max: 407: 113.94, s: 66.24

Time spent t 1

count: 353

min: 25, max: 425

: 115.99, s: 60.71

Table 1: Descriptive statistics for 400 deceptive opinion

spam submissions gathered using AMT. s corresponds to

the sample standard deviation.

indicates that any submission found to be of insuffi-

cient quality (e.g., written for the wrong hotel, unin-

telligible, unreasonably short,11

plagiarized,12

etc.)will be rejected.

It took approximately 14 days to collect 400 sat-

isfactory deceptive opinions. Descriptive statistics

appear in Table 1. Submissions vary quite dramati-

cally both in length, and time spent on the task. Par-

ticularly, nearly 12% of the submissions were com-

pleted in under one minute. Surprisingly, an inde-

pendent two-tailed t-test between the mean length of

these submissions (t


5/11

TRUTHFUL DECEPTIVE

Accuracy P R F P R F

HUMAN

JUDGE 1 61.9% 57.9 87.5 69.7 74.4 36.3 48.7

JUDGE 2 56.9% 53.9 95.0 68.8 78.9 18.8 30.3

JUDGE 3 53.1% 52.3 70.0 59.9 54.7 36.3 43.6

METAMAJORITY 58.1% 54.8 92.5 68.8 76.0 23.8 36.2

SKEPTIC 60.6% 60.8 60.0 60.4 60.5 61.3 60.9

Table 2: Performance of three human judges and two meta-judges on a subset of 160 opinions, corresponding to thefirst fold of our cross-validation experiments in Section 5. Boldface indicates the largest value for each column.

experiments described in Section 5, contains all 40

reviews from each of four randomly chosen hotels.

Unlike the Turkers, our student volunteers are not

offered a monetary reward. Consequently, we con-

sider their judgements to be more honest than those

obtained via AMT.

Additionally, to test the extent to which the in-

dividual human judges are biased, we evaluate the

performance of two virtual meta-judges. Specifi-cally, the MAJORITY meta-judge predicts decep-

tive when at least two out of three human judges

believe the review to be deceptive, and the SKEP-

TI C meta-judge predicts deceptive when any hu-

man judge believes the review to be deceptive.

Human and meta-judge performance is given in

Table 2. It is clear from the results that human

judges are not particularly effective at this task. In-

deed, a two-tailed binomial test fails to reject the

null hypothesis that JUDGE 2 and JUDGE 3 per-

form at-chance (p = 0.003, 0.10, 0.48 for the threejudges, respectively). Furthermore, all three judges

suffer from truth-bias (Vrij, 2008), a common find-

ing in deception detection research in which hu-

man judges are more likely to classify an opinion

as truthful than deceptive. In fact, JUDGE 2 clas-

sified fewer than 12% of the opinions as decep-

tive! Interestingly, this bias is effectively smoothed

by the SKEPTIC meta-judge, which produces nearly

perfectly class-balanced predictions. A subsequent

reevaluation of human performance on this task sug-

gests that the truth-bias can be reduced if judges

are given the class-proportions in advance, although

such prior knowledge is unrealistic; and ultimately,

performance remains similar to that of Table 2.

Inter-annotator agreement among the three

judges, computed using Fleiss kappa, is 0.11.

While there is no precise rule for interpreting

kappa scores, Landis and Koch (1977) suggest

that scores in the range (0.00, 0.20] correspond

to slight agreement between annotators. The

largest pairwise Cohens kappa is 0.12, between

JUDGE 2 and JUDGE 3a value far below generally

accepted pairwise agreement levels. We suspect

that agreement among our human judges is so

low precisely because humans are poor judges of

deception (Vrij, 2008), and therefore they perform

nearly at-chance respective to one another.

4 Automated Approaches to Deceptive

Opinion Spam Detection

We consider three automated approaches to detect-

ing deceptive opinion spam, each of which utilizes

classifiers (described in Section 4.4) trained on the

dataset of Section 3. The features employed by each

strategy are outlined here.

4.1 Genre identification

Work in computational linguistics has shown thatthe frequency distribution of part-of-speech (POS)

tags in a text is often dependent on the genre of the

text (Biber et al., 1999; Rayson et al., 2001). In our

genre identification approach to deceptive opinion

spam detection, we test if such a relationship exists

for truthful and deceptive reviews by constructing,

for each review, features based on the frequencies of

each POS tag.15 These features are also intended to

provide a good baseline with which to compare our

other automated approaches.

4.2 Psycholinguistic deception detection

The Linguistic Inquiry and Word Count (LIWC)

software (Pennebaker et al., 2007) is a popular au-

tomated text analysis tool used widely in the so-

cial sciences. It has been used to detect personality

15We use the Stanford Parser (Klein and Manning, 2003) to

obtain the relative POS frequencies.

313


6/11

traits (Mairesse et al., 2007), to study tutoring dy-

namics (Cade et al., 2010), and, most relevantly, to

analyze deception (Hancock et al., 2008; Mihalcea

and Strapparava, 2009; Vrij et al., 2007).

While LIWC does not include a text classifier, we

can create one with features derived from the LIWC

output. In particular, LIWC counts and groups

the number of instances of nearly 4,500 keywords

into 80 psychologically meaningful dimensions. We

construct one feature for each of the 80 LIWC di-

mensions, which can be summarized broadly under

the following four categories:

1. Linguistic processes: Functional aspects of text

(e.g., the average number of words per sen-

tence, the rate of misspelling, swearing, etc.)

2. Psychological processes: Includes all social,

emotional, cognitive, perceptual and biological

processes, as well as anything related to time orspace.

3. Personal concerns: Any references to work,

leisure, money, religion, etc.

4. Spoken categories: Primarily filler and agree-

ment words.

While other features have been considered in past

deception detection work, notably those of Zhou et

al. (2004), early experiments found LIWC features

to perform best. Indeed, the LIWC2007 software

used in our experiments subsumes most of the fea-tures introduced in other work. Thus, we focus our

psycholinguistic approach to deception detection on

LIWC-based features.

4.3 Text categorization

In contrast to the other strategies just discussed,

our text categorization approach to deception de-

tection allows us to model both content and con-

text with n-gram features. Specifically, we consider

the following three n-gram feature sets, with the

corresponding features lowercased and unstemmed:

UNIGRAMS, BIGRAMS+, TRIGRAMS+, where the

superscript + indicates that the feature set subsumes

the preceding feature set.

4.4 Classifiers

Features from the three approaches just introduced

are used to train Nave Bayes and Support Vector

Machine classifiers, both of which have performed

well in related work (Jindal and Liu, 2008; Mihalcea

and Strapparava, 2009; Zhou et al., 2008).

For a document x, with label y, the Nave Bayes

(NB) classifier gives us the following decision rule:

y = arg maxc

Pr(y = c) Pr(x | y = c) (1)

When the class prior is uniform, for example

when the classes are balanced (as in our case), (1)

can be simplified to the maximum likelihood classi-

fier (Peng and Schuurmans, 2003):

y = arg maxc

Pr(x | y = c) (2)

Under (2), both the NB classifier used by Mihal-

cea and Strapparava (2009) and the language model

classifier used by Zhou et al. (2008) are equivalent.Thus, following Zhou et al. (2008), we use the SRI

Language Modeling Toolkit (Stolcke, 2002) to esti-

mate individual language models, Pr(x | y = c),for truthful and deceptive opinions. We consider

all three n-gram feature sets, namely UNIGRAMS,

BIGRAMS+, and TRIGRAMS+, with corresponding

language models smoothed using the interpolated

Kneser-Ney method (Chen and Goodman, 1996).

We also train Support Vector Machine (SVM)

classifiers, which find a high-dimensional separating

hyperplane between two groups of data. To simplifyfeature analysis in Section 5, we restrict our evalu-

ation to linear SVMs, which learn a weight vector

w and bias term b, such that a document x can be

classified by:

y = sign(w x + b) (3)

We use SVMlight (Joachims, 1999) to train our

linear SVM models on all three approaches and

feature sets described above, namely PO S, LIWC,

UNIGRAMS, BIGRAMS+, and TRIGRAMS+. We also

evaluate every combination of these features, but

for brevity include only LIWC+BIGRAMS+, which

performs best. Following standard practice, doc-

ument vectors are normalized to unit-length. For

LIWC+BIGRAMS+, we unit-length normalize LIWC

and BIGRAMS+ features individually before com-

bining them.

314


7/11

TRUTHFUL DECEPTIVE

Approach Features Accuracy P R F P R F

G EN RE ID EN TIFIC AT ION PO SSVM 73.0% 75.3 68.5 71.7 71.1 77.5 74.2

PSYCHOLINGUISTICLIWCSVM 76.8% 77.2 76.0 76.6 76.4 77.5 76.9

DECEPTION DETECTION

TEXT CATEGORIZATION

UNIGRAMS SVM 88.4% 89.9 86.5 88.2 87.0 90.3 88.6

BIGRAMS+SVM 89.6% 90.1 89.0 89.6 89.1 90.3 89.7

LIWC+BIGRAMS+SVM 89.8% 89.8 89.8 89.8 89.8 89.8 89.8

TRIGRAMS

+

SVM 89.0% 89.0 89.0 89.0 89.0 89.0 89.0UNIGRAMS NB 88.4% 92.5 83.5 87.8 85.0 93.3 88.9

BIGRAMS+NB 88.9% 89.8 87.8 88.7 88.0 90.0 89.0

TRIGRAMS+NB 87.6% 87.7 87.5 87.6 87.5 87.8 87.6

HUMAN / META

JUDGE 1 61.9% 57.9 87.5 69.7 74.4 36.3 48.7

JUDGE 2 56.9% 53.9 95.0 68.8 78.9 18.8 30.3

SKEPTIC 60.6% 60.8 60.0 60.4 60.5 61.3 60.9

Table 3: Automated classifier performance for three approaches based on nested 5-fold cross-validation experiments.

Reported precision, recall and F-score are computed using a micro-average, i.e., from the aggregate true positive, false

positive and false negative rates, as suggested by Forman and Scholz (2009). Human performance is repeated here for

JUDGE 1, JUDGE 2 and the SKEPTIC meta-judge, although they cannot be directly compared since the 160-opinion

subset on which they are assessed only corresponds to the first cross-validation fold.

5 Results and Discussion

The deception detection strategies described in Sec-

tion 4 are evaluated using a 5-fold nested cross-

validation (CV) procedure (Quadrianto et al., 2009),

where model parameters are selected for each test

fold based on standardCV experiments on the train-

ing folds. Folds are selected so that each contains all

reviews from four hotels; thus, learned models are

always evaluated on reviews from unseen hotels.

Results appear in Table 3. We observe that auto-mated classifiers outperform human judges for every

metric, except truthful recall where JUDGE 2 per-

forms best.16 However, this is expected given that

untrained humans often focus on unreliable cues to

deception (Vrij, 2008). For example, one study ex-

amining deception in online dating found that hu-

mans perform at-chance detecting deceptive pro-

files because they rely on text-based cues that are

unrelated to deception, such as second-person pro-

nouns (Toma and Hancock, In Press).

Among the automated classifiers, baseline per-

formance is given by the simple genre identifica-

tion approach (PO SSV M) proposed in Section 4.1.

Surprisingly, we find that even this simple auto-

16As mentioned in Section 3.3, JUDGE 2 classified fewer than

12% of opinions as deceptive. While achieving 95% truthful re-

call, this judges corresponding precision was not significantly

better than chance (two-tailed binomial p = 0.4).

mated classifier outperforms most human judges

(one-tailed sign test p = 0.06, 0.01, 0.001 for thethree judges, respectively, on the first fold). This

result is best explained by theories of reality mon-

itoring (Johnson and Raye, 1981), which suggest

that truthful and deceptive opinions might be clas-

sified into informative and imaginative genres, re-

spectively. Work by Rayson et al. (2001) has found

strong distributional differences between informa-

tive and imaginative writing, namely that the former

typically consists of more nouns, adjectives, prepo-sitions, determiners, and coordinating conjunctions,

while the latter consists of more verbs,17 adverbs,18

pronouns, and pre-determiners. Indeed, we find that

the weights learned by PO SSV M (found in Table 4)

are largely in agreement with these findings, no-

tably except for adjective and adverb superlatives,

the latter of which was found to be an exception by

Rayson et al. (2001). However, that deceptive opin-

ions contain more superlatives is not unexpected,

since deceptive writing (but not necessarily imagi-

native writing in general) often contains exaggeratedlanguage (Buller and Burgoon, 1996; Hancock et al.,

2008).

Both remaining automated approaches to detect-

ing deceptive opinion spam outperform the simple

17Past participle verbs were an exception.18Superlative adverbs were an exception.

315


8/11

TRUTHFUL/INFORMATIVE DECEPTIVE/IMAGINATIVE

Category Variant Weight Category Variant Weight

NOUNS

Singular 0.008

VERBS

Base -0.057

Plural 0.002 Past tense 0.041

Proper, singular -0.041 Present participle -0.089

Proper, plural 0.091 Singular, present -0.031

ADJECTIVES

General 0.002 Third person0.026

Comparative 0.058 singular, present

Superlative -0.164 Modal -0.063PREPOSITIONS General 0.064ADVERBS

General 0.001

DETERMINERS General 0.009 Comparative -0.035

COORD. CONJ. General 0.094PRONOUNS

Personal -0.098

VERBS Past participle 0.053 Possessive -0.303

ADVERBS Superlative -0.094 PR E-DETERMINERS General 0.017

Table 4: Average feature weights learned by PO SSVM . Based on work by Rayson et al. (2001), we expect weights on

the left to be positive (predictive oftruthful opinions), and weights on the right to be negative (predictive of deceptive

opinions). Boldface entries are at odds with these expectations. We report average feature weights ofunit-normalized

weight vectors, rather than raw weights vectors, to account for potential differences in magnitude between the folds.

genre identification baseline just discussed. Specifi-

cally, the psycholinguistic approach (LIWCSV M) pro-

posed in Section 4.2 performs 3.8% more accurately

(one-tailed sign test p = 0.02), and the standard textcategorization approach proposed in Section 4.3 per-

forms between 14.6% and 16.6% more accurately.

However, best performance overall is achieved by

combining features from these two approaches. Par-

ticularly, the combined model LIWC+BIGRAMS+SV Mis 89.8% accurate at detecting deceptive opinion

spam.19

Surprisingly, models trained only on

UNIGRAMSthe simplest n-gram feature setoutperform all nontext-categorization approaches,

and models trained on BIGRAMS+ perform even

better (one-tailed sign test p = 0.07). This suggeststhat a universal set of keyword-based deception

cues (e.g., LIWC) is not the best approach to de-

tecting deception, and a context-sensitive approach

(e.g., BIGRAMS+) might be necessary to achieve

state-of-the-art deception detection performance.

To better understand the models learned by these

automated approaches, we report in Table 5 the top

15 highest weighted features for each class (truthfuland deceptive) as learned by LIWC+BIGRAMS+SV Mand LIWCSVM . In agreement with theories of reality

monitoring (Johnson and Raye, 1981), we observe

that truthful opinions tend to include more sensorial

and concrete language than deceptive opinions; in

19The result is not significantly better than BIGRAMS+SVM .

LIWC+BIGRAMS+SVM LIWCSVM

T RU TH FU L D EC EP TI VE T RU TH FU L D EC EP TI VE

- chicago hear i

... my number family

on hotel allpunct perspron

location , and negemo see

) luxury dash pronoun

allpunctLIWC experience exclusive leisure

floor hilton we exclampunct

( business sexual sixletters

the hotel vacation period posemo

bathroom i otherpunct comma

small spa space cause

helpful looking human auxverb

$ while past future

hotel . husband inhibition perceptualother my husband assent feel

Table 5: Top 15 highest weighted truthful and deceptive

features learned by LIWC+BIGRAMS+SVM and LIWCSVM .

Ambiguous features are subscripted to indicate the source

of the feature. LIWC features correspond to groups

of keywords as explained in Section 4.2; more details

about LIWC and the LIWC categories are available athttp://liwc.net.

particular, truthful opinions are more specific about

spatial configurations (e.g., small, bathroom, on, lo-

cation). This finding is also supported by recent

work by Vrij et al. (2009) suggesting that liars have

considerable difficultly encoding spatial information

into their lies. Accordingly, we observe an increased

focus in deceptive opinions on aspects external to

the hotel being reviewed (e.g., husband, business,

316


9/11

vacation).

We also acknowledge several findings that, on the

surface, are in contrast to previous psycholinguistic

studies of deception (Hancock et al., 2008; Newman

et al., 2003). For instance, while deception is often

associated with negative emotion terms, our decep-

tive reviews have more positive and fewer negative

emotion terms. This pattern makes sense when one

considers the goal of our deceivers, namely to create

a positive review (Buller and Burgoon, 1996).

Deception has also previously been associated

with decreased usage of first person singular, an ef-

fect attributed to psychological distancing (Newman

et al., 2003). In contrast, we find increased first

person singular to be among the largest indicators

of deception, which we speculate is due to our de-

ceivers attempting to enhance the credibility of their

reviews by emphasizing their own presence in the

review. Additional work is required, but these find-ings further suggest the importance of moving be-

yond a universal set of deceptive language features

(e.g., LIWC) by considering both the contextual (e.g.,

BIGRAMS+) and motivational parameters underly-

ing a deception as well.

6 Conclusion and Future Work

In this work we have developed the first large-scale

dataset containing gold-standard deceptive opinion

spam. With it, we have shown that the detectionof deceptive opinion spam is well beyond the ca-

pabilities of human judges, most of whom perform

roughly at-chance. Accordingly, we have introduced

three automated approaches to deceptive opinion

spam detection, based on insights coming from re-

search in computational linguistics and psychology.

We find that while standard n-grambased text cate-

gorization is the best individual detection approach,

a combination approach using psycholinguistically-

motivated features and n-gram features can perform

slightly better.

Finally, we have made several theoretical con-

tributions. Specifically, our findings suggest the

importance of considering both the context (e.g.,

BIGRAMS+) and motivations underlying a decep-

tion, rather than strictly adhering to a universal set

of deception cues (e.g., LIWC). We have also pre-

sented results based on the feature weights learned

by our classifiers that illustrate the difficulties faced

by liars in encoding spatial information. Lastly, we

have discovered a plausible relationship between de-

ceptive opinion spam and imaginative writing, based

on POS distributional similarities.

Possible directions for future work include an ex-

tended evaluation of the methods proposed in this

work to both negative opinions, as well as opinions

coming from other domains. Many additional ap-

proaches to detecting deceptive opinion spam are

also possible, and a focus on approaches with high

deceptive precision might be useful for production

environments.

Acknowledgments

This work was supported in part by National

Science Foundation Grants BCS-0624277, BCS-

0904822, HSD-0624267, IIS-0968450, and NSCC-

0904822, as well as a gift from Google, and the

Jack Kent Cooke Foundation. We also thank, al-

phabetically, Rachel Boochever, Cristian Danescu-

Niculescu-Mizil, Alicia Granstein, Ulrike Gretzel,

Danielle Kirshenblat, Lillian Lee, Bin Lu, Jack

Newton, Melissa Sackler, Mark Thomas, and Angie

Yoo, as well as members of the Cornell NLP sem-

inar group and the ACL reviewers for their insight-

ful comments, suggestions and advice on various as-

pects of this work.

References

C. Akkaya, A. Conrad, J. Wiebe, and R. Mihalcea. 2010.

Amazon mechanical turk for subjectivity word sense

disambiguation. In Proceedings of the NAACL HLT

2010 Workshop on Creating Speech and Language

Data with Amazons Mechanical Turk, Los Angeles,

pages 195203.

D. Biber, S. Johansson, G. Leech, S. Conrad, E. Finegan,

and R. Quirk. 1999. Longman grammar of spoken and

written English, volume 2. MIT Press.

C.F. Bond and B.M. DePaulo. 2006. Accuracy of de-

ception judgments. Personality and Social PsychologyReview, 10(3):214.

D.B. Buller and J.K. Burgoon. 1996. Interpersonal

deception theory. Communication Theory, 6(3):203

242.

W.L. Cade, B.A. Lehman, and A. Olney. 2010. An ex-

ploration of off topic conversation. In Human Lan-

guage Technologies: The 2010 Annual Conference of

317


10/11

the North American Chapter of the Association for

Computational Linguistics, pages 669672. Associa-

tion for Computational Linguistics.

S.F. Chen and J. Goodman. 1996. An empirical study of

smoothing techniques for language modeling. In Pro-

ceedings of the 34th annual meeting on Association

for Computational Linguistics, pages 310318. Asso-

ciation for Computational Linguistics.C. Danescu-Niculescu-Mizil, G. Kossinets, J. Kleinberg,

and L. Lee. 2009. How opinions are received by on-

line communities: a case study on amazon.com help-

fulness votes. In Proceedings of the 18th international

conference on World wide web, pages 141150. ACM.

H. Drucker, D. Wu, and V.N. Vapnik. 2002. Support

vector machines for spam categorization. Neural Net-

works, IEEE Transactions on, 10(5):10481054.

G. Forman and M. Scholz. 2009. Apples-to-Apples in

Cross-Validation Studies: Pitfalls in Classifier Perfor-

mance Measurement. ACM SIGKDD Explorations,

12(1):4957.

Z. Gyongyi, H. Garcia-Molina, and J. Pedersen. 2004.

Combating web spam with trustrank. In Proceedings

of the Thirtieth international conference on Very large

data bases-Volume 30, pages 576587. VLDB Endow-

ment.

J.T. Hancock, L.E. Curry, S. Goorha, and M. Woodworth.

2008. On lying and being lied to: A linguistic anal-

ysis of deception in computer-mediated communica-

tion. Discourse Processes, 45(1):123.

J. Jansen. 2010. Online product research. Pew Internet

& American Life Project Report.

N. Jindal and B. Liu. 2008. Opinion spam and analysis.

In Proceedings of the international conference on Web

search and web data mining, pages 219230. ACM.

T. Joachims. 1998. Text categorization with support vec-

tor machines: Learning with many relevant features.

Machine Learning: ECML-98, pages 137142.

T. Joachims. 1999. Making large-scale support vec-

tor machine learning practical. In Advances in kernel

methods, page 184. MIT Press.

M.K. Johnson and C.L. Raye. 1981. Reality monitoring.

Psychological Review, 88(1):6785.

S.M. Kim, P. Pantel, T. Chklovski, and M. Pennacchiotti.

2006. Automatically assessing review helpfulness.

In Proceedings of the 2006 Conference on EmpiricalMethods in Natural Language Processing, pages 423

430. Association for Computational Linguistics.

D. Klein and C.D. Manning. 2003. Accurate unlexical-

ized parsing. In Proceedings of the 41st Annual Meet-

ing on Association for Computational Linguistics-

Volume 1, pages 423430. Association for Computa-

tional Linguistics.

J.R. Landis and G.G. Koch. 1977. The measurement of

observer agreement for categorical data. Biometrics,

33(1):159.

E.P. Lim, V.A. Nguyen, N. Jindal, B. Liu, and H.W.

Lauw. 2010. Detecting product review spammers us-

ing rating behaviors. In Proceedings of the 19th ACM

international conference on Information and knowl-

edge management, pages 939948. ACM.S.W. Litvin, R.E. Goldsmith, and B. Pan. 2008. Elec-

tronic word-of-mouth in hospitality and tourism man-

agement. Tourism management, 29(3):458468.

F. Mairesse, M.A. Walker, M.R. Mehl, and R.K. Moore.

2007. Using linguistic cues for the automatic recogni-

tion of personality in conversation and text. Journal of

Artificial Intelligence Research, 30(1):457500.

R. Mihalcea and C. Strapparava. 2009. The lie detector:

Explorations in the automatic recognition of deceptive

language. In Proceedings of the ACL-IJCNLP 2009

Conference Short Papers, pages 309312. Association

for Computational Linguistics.

M.L. Newman, J.W. Pennebaker, D.S. Berry, and J.M.

Richards. 2003. Lying words: Predicting deception

from linguistic styles. Personality and Social Psychol-

ogy Bulletin, 29(5):665.

A. Ntoulas, M. Najork, M. Manasse, and D. Fetterly.

2006. Detecting spam web pages through content

analysis. In Proceedings of the 15th international con-

ference on World Wide Web, pages 8392. ACM.

M.P. OMahony and B. Smyth. 2009. Learning to rec-

ommend helpful hotel reviews. In Proceedings of

the third ACM conference on Recommender systems,

pages 305308. ACM.

F. Peng and D. Schuurmans. 2003. Combining naive

Bayes and n-gram language models for text classifica-

tion. Advances in Information Retrieval, pages 547

547.

J.W. Pennebaker, C.K. Chung, M. Ireland, A. Gonzales,

and R.J. Booth. 2007. The development and psycho-

metric properties of LIWC2007. Austin, TX, LIWC.

Net.

N. Quadrianto, A.J. Smola, T.S. Caetano, and Q.V.

Le. 2009. Estimating labels from label proportions.

The Journal of Machine Learning Research, 10:2349

2374.

P. Rayson, A. Wilson, and G. Leech. 2001. Grammaticalword class variation within the British National Cor-

pus sampler. Language and Computers, 36(1):295

306.

R.A. Rigby and D.M. Stasinopoulos. 2005. Generalized

additive models for location, scale and shape. Jour-

nal of the Royal Statistical Society: Series C (Applied

Statistics), 54(3):507554.

318


11/11

F. Sebastiani. 2002. Machine learning in automated

text categorization. ACM computing surveys (CSUR),

34(1):147.

M.A. Serrano, A. Flammini, and F. Menczer. 2009.

Modeling statistical properties of written text. PloS

one, 4(4):5372.

A. Stolcke. 2002. SRILM-an extensible language mod-

eling toolkit. In Seventh International Conference onSpoken Language Processing, volume 3, pages 901

904. Citeseer.

C. Toma and J.T. Hancock. In Press. What Lies Beneath:

The Linguistic Traces of Deception in Online Dating

Profiles. Journal of Communication.

A. Vrij, S. Mann, S. Kristen, and R.P. Fisher. 2007. Cues

to deception and ability to detect lies as a function

of police interview styles. Law and human behavior,

31(5):499518.

A. Vrij, S. Leal, P.A. Granhag, S. Mann, R.P. Fisher,

J. Hillman, and K. Sperry. 2009. Outsmarting the

liars: The benefit of asking unanticipated questions.

Law and human behavior, 33(2):159166.A. Vrij. 2008. Detecting lies and deceit: Pitfalls and

opportunities. Wiley-Interscience.

W. Weerkamp and M. De Rijke. 2008. Credibility im-

proves topical blog post retrieval. ACL-08: HLT,

pages 923931.

M. Weimer, I. Gurevych, and M. Muhlhauser. 2007. Au-

tomatically assessing the post quality in online discus-

sions on software. In Proceedings of the 45th An-

nual Meeting of the ACL on Interactive Poster and

Demonstration Sessions, pages 125128. Association

for Computational Linguistics.

G. Wu, D. Greene, B. Smyth, and P. Cunningham. 2010.

Distortion as a validation criterion in the identification

of suspicious reviews. Technical report, UCD-CSI-

2010-04, University College Dublin.

K.H. Yoo and U. Gretzel. 2009. Comparison of De-

ceptive and Truthful Travel Reviews. Information and

Communication Technologies in Tourism 2009, pages

3747.

L. Zhou, J.K. Burgoon, D.P. Twitchell, T. Qin, and J.F.

Nunamaker Jr. 2004. A comparison of classifica-

tion methods for predicting deception in computer-

mediated communication. Journal of Management In-

formation Systems, 20(4):139166.

L. Zhou, Y. Shi, and D. Zhang. 2008. A Statistical Lan-guage Modeling Approach to Online Deception De-

tection. IEEE Transactions on Knowledge and Data

Engineering, 20(8):10771081.

319

Finding Deceptive Opinion Spam by Any Stretch of the Imagination

Documents

Transcript of Finding Deceptive Opinion Spam by Any Stretch of the Imagination